STA218 Analysis of Variance

Transcription

1 STA218 Analysis of Variance Al Nosedal. University of Toronto. Fall 2017 November 27, 2017

2 The Data Matrix The following table shows last year s sales data for a small business. The sample is put into a matrix format in which each of the three rows corresponds to one of the three countries in which the company does business, and each of the four columns corresponds to one of its four salespersons. So a cell in the matrix corresponds to one of 12 salesperson/country combinations. The numbers in the cell represent the sales (in units of $1000) made by that particular salesman in that country last year. This data will be used throughout the chapter to develop the theory underlying Analysis of Variance or, for short, ANOVA.

3 The Data Matrix Country A Country B Country C Average Salesperson 1 6, 7, 8 10, 10 12, 13, Salesperson 2 10, 15 10, 10 11, Salesperson 3 10, 15 7, 13 11, Salesperson 4 2, 3, 4 10, 10 8, 9, 10 7 Average

4 Altogether, there were 28 sales last year that totaled $280 - so the average sale was $10. The row (salesperson) averages are: Row 1 (Salesperson 1) $10. Row 2 (Salesperson 2) $12. Row 3 (Salesperson 3) $12. Row 4 (Salesperson 4) $ 7.

5 The column (country) averages are: Column 1 (Country A) $8. Column 2 (Country B) $10. Column 3 (Country C) $12.

6 Now we will begin our study of how to make a statistically valid prediction of the next sales figure. In that regard, there are four possible situations that can occur. 1. Neither the country nor the salesperson of the next sale (observation) is known. 2. The country of the next sale is known, but the salesperson is not known. 3. The salesperson of the next sale is known, but the country is not known. 4. Both the country and the salesperson of the next sale are known.

7 Situation 1. Without any additional information, the best prediction is the sample mean $10. This prediction is best in the least squares sense - that is, if $10 had been used to predict each of the 28 observations in the sample, then the total of the squared errors SS TOTAL would be as small as possible. In our data set, SS TOTAL equals 354. That figure can be verified by calculating (x i 10) 2 for each observation x i of the sample.

8 Situation 2. One-factor ANOVA Model. If only the country of the next sale is known, then two different predictions are possible for the next sales figure: The sample mean $10. The mean of the sales of the country in which the next sale will occur. (In this case, $8 if the next sale will occur in Country A, $10 in Country B, or $12 in Country C.) This prediction ignores the information present in the sales figures from the other two countries.

9 Situation 3. One-factor ANOVA Model. If only the salesperson of the next sale is known, then two different predictions are possible for the next sales figure: The sample mean $10. The mean of the previous sales of the salesperson who will make the next sale. (In this case, $10 if the next sale will be made by Salesperson 1, etc. ) This prediction ignores the information present in the sales figures from the other three salespersons.

10 Situations 2 and 3 are called one-factor ANOVA models, because only one-factor is known about the next sale. As noted, if a prediction is either a row mean or a column mean, then it ignores the observations in the other rows or columns. To make that kind of prediction, it s necessary to statistically verify that the ignored observations are indeed different populations and therefore not relevant to the prediction.

11 Situation 4. Two-factor ANOVA Model. We are not covering this kind of model in our course.

12 The Null Hypothesis for One-Factor ANOVA We have discussed the prediction possibilities for one-factor ANOVA models. Now, we will learn how to test the statistical significance of a one-factor ANOVA model. Let s suppose that we want to predict the next sales figure, and that we know the country in which this sale will occur but the identity of the salesperson is NOT known. Without any statistical testing, we can always by default use the sample mean $10 to predict the next sale. The default prediction, the sample mean, doesn t use any information about the country (column) in which the sale will occur.

13 The Null Hypothesis for One-Factor ANOVA However, if instead we use the mean of the observations in only one column (the column that corresponds to the particular country in which we know the next sale will occur), then we have to test the null hypothesis H 0 : µ COL1, µ COL2, µ COL3, are equal and reject it in favor of the alternative hypothesis H a : µ COL1, µ COL2, µ COL3, are NOT all equal

14 The Null Hypothesis for One-Factor ANOVA If the null hypothesis is rejected, then we can be statistically confident that the column means are not all equal, and therefore that the individual column means (i. e., $8, $10, $12) can be used to predict the amount of the next sale. If the next sale sale was going to occur in Country A, then the prediction would be $8. If the next sale was going to occur in Country B, then the prediction would be $10. If the next sale was going to occur in Country C, then the prediction would be $12.

15 The One-Factor ANOVA F Test To test the null hypothesis stated above, we have to calculate an F-statistic. If F STAT > F (c 1,n c), α, then reject H 0, and use the sample column means to predict future observations. Otherwise, do not reject H 0 and use the overall sample mean to predict future observations.

16 ANOVA Table To see how this F STAT is calculated, see the ANOVA Table below. Source of Degrees of Sum of Mean Sum of F Ratio Variation Freedom Square Squares (df) (SS) (MSS) SS Explained c -1 SS EXP EXP c 1 F = MSS EXP MSS UNEXP SS Unexplained n -c SS UNEXP UNEXP n c Total n -1 SS TOTAL

17 Calculation of SS TOTAL If no model is used, then the predictions for each of the 28 observations (in dollar amounts) will be 10. If these predictions are used, the squared error of these 28 predictions is given in the table below. Country A Country B Country C Salesperson 1 16, 9, 4 0, 0 4, 9, 16 Salesperson 2 0, 25 0, 0 1, 36 Salesperson 3 0, 25 9, 9 1, 36 Salesperson 4 64, 49, 36 0, 0 4, 1, 0 Prediction Errors Squared when NO Factor is used (Total) = 354.

18 Calculation of SS UNEXPLAINED If the column model is used, then the 28 observations would have the following 28 predictions, where $8 is the average for the first column, $10 is the average for the second column, and $12 is the average for the third column. Country A Country B Country C Salesperson 1 8, 8, 8 10, 10 12, 12, 12 Salesperson 2 8, 8 10, 10 12, 12 Salesperson 3 8, 8 10, 10 12, 12 Salesperson 4 8, 8, 8 10, 10 12, 12, 12

19 Calculation of SS UNEXPLAINED Using the above 28 predictions, the errors squared are shown in the table below. Country A Country B Country C Salesperson 1 4, 1, 0 0, 0 0, 1, 4 Salesperson 2 4, 49 0, 0 1, 16 Salesperson 3 4, 49 9, 9 1, 16 Salesperson 4 36, 25, 16 0, 0 16, 9, 4 Errors Squared when the Column Factor is used (Total) = 274.

20 Calculation of SS EXPLAINED The units explained by the column model are calculated by finding the square of each prediction change when moving from NO model to the column model. The following table presents the square of each prediction change: Country A Country B Country C Salesperson 1 4, 4, 4 0, 0 4, 4, 4 Salesperson 2 4, 4 0, 0 4, 4 Salesperson 3 4, 4 0, 0 4, 4 Salesperson 4 4, 4, 4 0, 0 4, 4, 4 Table of the Square of the Prediction Change when Moving from NO Model to the Column Model (Total) = 80.

21 ANOVA Table The ANOVA Table for the column factor can now be filled in as shown below: Source of Degrees of Sum of Mean Sum of F Ratio Variation Freedom Square Squares (df) (SS) (MSS) 80 Explained = = Unexplained = Total So for this one-factor ANOVA model, F STAT = 3.65.

22 Conclusion If the null hypothesis is true, then the F-statistic should be a value from the F 2, 25 distribution. Referring to the table that contains the upper 0.05 cut-off points of F distributions, we see that F (2,25),0.05 = Since 3.65 is greater than 3.39, this tells us that the F-statistic is in the upper 0.05 of the F 2, 25 distribution. Therefore we can reject the null hypothesis at the 0.05 significance level, and we conclude that the country means are not all the same. Thus, the prediction for the next sale in a known country is the mean of all the previous sales in that country.

23 This time for the Row Factor We have just performed the F test to verify that the country (column) one-factor ANOVA model is statistically significant. There is another one-factor ANOVA model that also could be examined - the salesperson (row) factor model. Let s test H 0 : µ ROW 1, µ ROW 2, µ ROW 3, µ ROW 4 and reject it in favor of the alternative hypothesis are equal H a : µ ROW 1, µ ROW 2, µ ROW 3, µ ROW 4 are NOT all equal

24 ANOVA Table The resulting ANOVA table for the salesperson (row) factor is shown below: Source of Degrees of Sum of Mean Sum of F Ratio Variation Freedom Square Squares (df) (SS) (MSS) 120 Explained = = Unexplained = 9.75 Total So for this one-factor ANOVA model, F STAT = 4.10.

25 Conclusion Consulting the upper 0.05 cut-off table for the F distribution, we find that F (3, 24),0.05 = 3.01 Since F-statistic = 4.10 > 3.01, the null hypothesis can once again be rejected at the 0.05 level, and we can use the salesperson factor to predict sales, concluding that it is statistically valid to predict either $10, $12, or $7, respectively, for Salespersons 1, 2, 3, or 4.

26 # R Code; sales1=c(6, 7, 8, 10, 10, 12, 13, 14); sales2=c (10, 15, 10, 10, 11, 16 ); sales3=c(10, 15, 7, 13, 11, 16 ); sales4= c(2, 3, 4, 10, 10, 8, 9, 10 ); sales=c(sales1,sales2,sales3,sales4); person=c(rep(1,8),rep(2,6),rep(3,6),rep(4,8)); oneway.test(sales~person,var.equal=true);

27 ## ## One-way analysis of means ## ## data: sales and person ## F = , num df = 3, denom df = 24, p-value =

28 Underlying Assumptions Officially, to use the predictions from an ANOVA model, three assumptions about the populations from which the sample was taken must be satisfied: 1. Each population has a Normal distribution. 2. Each population has the same standard deviation σ. 3. The observations are mutually independent of one another.

29 Formulas Sum of Squares for Treatments (a.k.a. between-treatments variation or Explained) SST = k n j ( x j x) 2 j=1 Sum of Squares for Error (a.k.a. within-treatments variation or Unexplained) SSE = n k j (x ij x j ) 2 = (n 1 1)s (n k 1)sk 2. j=1 i=1

30 Formulas Mean Square for Treatments Mean Square for Error MST = SST k 1 MSE = SSE n k

31 Formulas Test Statistic F = MST MSE

32 Exercise 14.1 A statistics practitioner calculated the following statistics: Treatment Statistic n x s Complete the ANOVA table.

33 Solution x = 5(10)+5(15)+5(20) = 15 SST = 5(10 15) 2 + 5(15 15) 2 + 5(20 15) 2 = 250 SSE = (5 1)(50) + (5 1)(50) + (5 1)(50) = 600

34 ANOVA Table Source of Degrees of Sum of Mean Sum of F Ratio Variation Freedom Square Squares (df) (SS) (MSS) 250 Treatments = = Error = 50 Total

35 Exercise 14.2 A statistics practitioner calculated the following statistics: Treatment Statistic n x s Complete the ANOVA table.

36 Solution x = 4(20)+4(22)+4(25) = SST = 4( ) 2 + 4( ) 2 + 5( ) 2 = SSE = (4 1)(10) + (4 1)(10) + (4 1)(10) = 90

37 ANOVA Table Source of Degrees of Sum of Mean Sum of F Ratio Variation Freedom Square Squares (df) (SS) (MSS) Treatments = = Error = 10 Total

38 Exercise 14.5 A consumer organization was concerned about the differences between the advertised sizes of containers and the actual amount of product. In a preliminary study, six packages of three different brands of margarine that are supposed to contain 500ml were measured. The differences from 500 ml are listed here. Do these data provide sufficient evidence to conclude that differences exist between the three brands? Use α = Brand 1 Brand 2 Brand

39 Solution Step 1. State Hypotheses. µ i = population mean for differences from 500 ml (brand i, where i = 1, 2, 3). H 0 : µ 1 = µ 2 = µ 3 H a : At least two means differ.

40 Solution Step 2. Compute test statistic. Brand 1 Brand 2 Brand 3 Mean Variance Grand mean = x = SST = 6( ) 2 + 6( ) 2 + 6( ) 2 = SSE = (6 1)(1.87) + (6 1)(2.30) + (6 1)(1.47) = 28.20

41 Solution Grand mean = x = SST = 6( ) 2 + 6( ) 2 + 6( ) 2 = SSE = (6 1)(1.87) + (6 1)(2.30) + (6 1)(1.47) = 28.20

42 ANOVA Table Source of Degrees of Sum of Mean Sum of F Ratio Variation Freedom Square Squares (df) (SS) (MSS) Treatments = = Error = 1.88 Total

43 Solution Step 3. Find Rejection Region. We reject the null hypothesis only if F > F α,k 1,n k If we let α = 0.05, the rejection region for this exercise is F > F 0.05, 2,15 = 3.682

44 Solution Step 4. Conclusion. We found the value of the test statistic to be F = Since F = 1.70 < F 0.05, 2,15 = 3.682, we can t reject H 0. Thus, there is not evidence to infer that the average differences differ between the three brands.

45 # R Code; brand1=c(1,3,3,0,1,0); brand2=c (2,2,4,3,0,4); brand3=c(1,2,4,2,3,4); differences=c(brand1,brand2,brand3); brand=c(rep(1,6),rep(2,6),rep(3,6)); oneway.test(differences~brand,var.equal=true);

46 ## ## One-way analysis of means ## ## data: differences and brand ## F = , num df = 2, denom df = 15, p-value =

47 Exercise The friendly folks a the Internal Revenue Service (IRS) in the United States and Canada Revenue Agency (CRA) are always looking for ways to improve the wording and format of its tax return forms. Three new forms have been developed recently. To determine which, if any, are superior to the current form, 120 individuals were asked to participate in an experiment. Each of the three new forms and the currently used form were filled out by 30 different people. The amount of time (in minutes) taken by each person to complete the task was recorded. What conclusions can be drawn from these data?

48 R Code #Step 1. Entering data; # importing data; # url of tax return forms; forms_url = " forms_data= read.table(forms_url,header=true); names(forms_data); forms_data[1:4, ];

49 R Code ## [1] "Form1" "Form2" "Form3" "Form4" ## Form1 Form2 Form3 Form4 ## ## ## ##

50 R Code #Step 2. ANOVA; time1=forms_data$form1; time2=forms_data$form2; time3=forms_data$form3; time4=forms_data$form4; length(forms_data$form1); times=c(time1,time2,time3,time4); forms=c(rep(1,30),rep(2,30),rep(3,30),rep(4,30)); oneway.test(times~forms,var.equal=true)

51 R Code ## [1] 30 ## ## One-way analysis of means ## ## data: times and forms ## F = , num df = 3, denom df = 116, p-value =